bonlime

@ternaus I don't think it's useful to add something that could be achieved by 2 lines of current augs. For me augmentations list is always pretty long and has 10+...

@Moxinilian you're right. If you're interested in more efficient implementation you could check TF BatchNorm + `virtual_batch_size` param. They reshape the input and then batch norm it __inside__ the BN...

i know it's very late, but the problem is - you have to normalize self.weight by different dimension in order to obtain real cosine: `cosine = F.linear(F.normalize(input), F.normalize(self.weight, dim=0))`

I don't think you can make this script compatible with TF 2.0. They made a really huge change, now there is no such thing as session and there is no...

It probably works for authors because they add Shift-Conv (which is almost a conv3x3), default SwinIR doesn't have any 3x3 convs so postitional embedding is required

On my 1080ti it is possible to have a batch of 6 images with size (512, 512) and OS 16 don't know what's wrong with your code.

just initialize new model with `OS=8` and load weights from model trained on `OS=16`. That should work

@Meight You raised a very good point. After implementing this model I also tried very hard to fine-tune it, but the results were unsatisfactory bad. I stopped trying at the...

I've managed to successfully fine-tune models from this repo: https://github.com/qubvel/segmentation_models, maybe you can use them as well.

Hi, the Cityscapes weights were not trained, they were transferred from TF version of Deeplab. I haven't tested the weights myself, but you could test them and report the results...